CN115792677A

CN115792677A - Lithium ion battery life prediction method based on improved ELM

Info

Publication number: CN115792677A
Application number: CN202211526361.0A
Authority: CN
Inventors: 袁慧梅; 唐婷; 冉波
Original assignee: Capital Normal University
Current assignee: Capital Normal University
Priority date: 2020-12-21
Filing date: 2020-12-21
Publication date: 2023-03-14
Also published as: CN112731183B; CN112731183A

Abstract

The invention discloses a lithium ion battery service life prediction method based on improved ELM, wherein the full connection relation between an input layer and a hidden layer of the ELM is changed into convolution operation, namely, a convolution kernel with a conventional size is introduced, the convolution kernel is regarded as the deformation of the connection weight between the original input layer and the hidden layer and is convolved with the data of the input layer, an extracted characteristic matrix is subjected to average pooling to obtain a hidden layer output matrix H, and Moore-Penrose generalized inverse H of H is further subjected to average pooling to obtain a hidden layer output matrix H ⁺ And multiplying the output weight matrix beta by the output matrix of the training set to obtain an output weight matrix beta, and substituting the output weight matrix beta into the data of the test set to predict the result. The other is to change the full connection relation between the ELM input layer and the hidden layer into pooling, i.e. the input layer data is directly obtained by poolingThe hidden layer outputs a matrix H, and then Moore-Penrose generalized inverse H of the H ⁺ And multiplying the output weight matrix beta by the output matrix of the training set to obtain an output weight matrix beta, and substituting the output weight matrix beta into the data of the test set to predict the result. The invention ensures that the prediction result is more accurate and the robustness is stronger.

Description

Lithium ion battery life prediction method based on improved ELM

The application is a divisional application with the application date of 2020, 12 and 21, and the application number of 202011518358.5, namely 'a lithium ion battery life prediction method based on improved ELM'.

Technical Field

The invention relates to a lithium ion battery service life prediction method based on an improved ELM (extreme learning machine algorithm), and belongs to the technical field of lithium battery health management.

Background

The lithium battery has the advantages of small volume, light weight, high energy density, wide working range, high charge and discharge rate, environmental protection and the like, and can be widely applied to various industries, for example: mobile phones, computers, airplanes, power tools, automobiles, cameras, and the like. Therefore, the lithium battery can be seen everywhere in daily life, and the application prospect of the lithium battery is very considerable. However, due to various external environmental factors and internal factors, the lithium battery inevitably undergoes an aging phenomenon after being used for a certain period of time, significantly affecting the performance thereof. In severe cases, when the performance of a lithium battery is reduced to a certain extent, the problems of battery leakage, insulation damage and partial short circuit can cause catastrophic accidents. Therefore, the subjects of Prognosis and Health Management (PHM) are important. Residual Useful Life (RUL) is an important research direction for this subject, and its accurate prediction plays an important role in safety, reliability and economy.

Generally, the RUL of a lithium battery is defined as the number of remaining charge and discharge cycles before the performance degrades to a rated failure threshold. The degradation of lithium batteries can be characterized by Health Indicators (HIs), such as current, voltage, impedance and capacity, wherein especially the capacity degradation and RUL prediction of lithium batteries have become a matter of great concern. In the existing literature, the prediction of RUL for lithium batteries can be roughly divided into three categories: model-based methods, data-driven methods, and hybrid methods.

In the model-based method, an electrochemical model, an equivalent circuit model, or an empirical model may be used as the model. The electrochemical model is constructed according to a specific physicochemical reaction process inside the battery, and can describe the intrinsic cause and complex mechanism of the battery aging in detail and accurately. Although the model deeply understands the essential characteristics of the battery and the well-established model can well predict the RUL, many internal factors and various external interferences need to be considered in the process of establishing the model, so that a large number of parameters in the model need to be estimated, the calculation cost is quite high, and the model is difficult to realize in practical application. The equivalent circuit model is to simplify circuit analysis, and a simple component circuit is constructed by using resistance elements such as a resistance capacitor and a voltage source, so as to approximately simulate the dynamic characteristics of a battery. Common equivalent circuit models of lithium batteries include: rint model, RC model, thevenin model, PNGV model and the like, and meanwhile, the derived multiple model parameter identification methods further modify the models, so that the models have high reliability. This model is more suitable than an electrochemical model, but involves a large number of parameter estimates as well, and is still quite complex. Unlike these two models, an empirical model is obtained by using parameters that reflect the performance of the battery, such as: the capacity, the internal resistance, the discharge final voltage and the like are obtained, so that the change trend of the state parameters along with the time or the recursion relation between the state variables at the front moment and the rear moment of the system is established, and the prediction of the RUL of the battery is realized. He et al proposed a four-parameter empirical exponential model of capacity degradation, which was curve-fitted to the capacity fade data of the cell. Based on this model, many filtering algorithms such as: extended Kalman Filter (EKF), unscented Kalman Filter (UKF), particle Filter (PF), unscented Particle Filter (UPF), etc. are widely used for prediction of RUL. Xing et al propose a combined model to describe the capacity drop, which combines an empirical index and a polynomial regression model, and then uses PF to adjust the model parameters on-line, thereby realizing the prediction of the RUL of the lithium battery. Miao et al proposed an improved UPF method based on MCMC to predict the RUL of lithium batteries. Duong et al introduced a Heuristic Kalman Algorithm (HKA), which combined with PF could predict the RUL of lithium batteries with high accuracy. It follows that the improvement of this process by the extensive literature is mainly focused on two points: the method comprises the following steps of firstly, improving an empirical model and secondly, improving a filtering algorithm. In addition to capacity-based empirical models, there are other types of empirical models, such as resistance-based exponential growth models and linear parameter variation models. Compared with the former two models, the empirical model of the lithium battery is easier to construct, the workload is greatly reduced, the application range is wider, but the empirical model is easily interfered by external environmental conditions such as noise, and the adaptability and the robustness are poorer. In addition, particle depletion and initialization of parameters can present problems that make it difficult to accurately construct the model.

The prediction method based on data driving is based on historical working data, performance degradation data of the lithium battery are analyzed and mined through various data processing methods, and then corresponding relations are established according to the data so as to realize RUL prediction. Commonly used data driving methods mainly include Support Vector Machine (SVM), relevance Vector Machine (RVM), autoregressive Integrated Moving Average model (ARIMA), hidden Markov Model (HMM), neural Network (NN) method, and the like. Neural networks are an important branch of the field of artificial intelligence and are of great interest in data-driven prediction of RUL. Wu et al uses a Feed Forward Neural Network (FFNN) to simulate the relationship between a battery charging curve and a battery RUL under constant current, and in consideration of the nonlinearity of the curve, an input sampling point of the FFNN IS obtained from Importance Sampling (IS), so that the on-line estimation of the lithium battery RUL IS realized. Xing et al established a relationship between battery impedance and capacity using an adaptive neural network with battery impedance as input data and battery capacity as output data, resulting in a more accurate and competitive prediction result. Liu et al propose a self-Adaptive Recurrent Neural Network (ARNN) for predicting the remaining life of a lithium ion battery, which makes full use of the previous state of the system through self-adaptation and recursive feedback, improving the prediction accuracy, and the experimental results are superior to the classical Recurrent Neural Network (RNN) and the Recurrent Neural Fuzzy System (RNFs). Since RNN cannot handle "Long-Term dependencies," the Long-Short Term Memory (LSTM) network proposed by Hochreiter et al has become the focus of RUL prediction research. Zhang et al learns the long-term dependence between the capacity drops of the lithium batteries by using LSTM RNNs, and performs adaptive optimization on the constructed neural network by using a Root Mean Square backward propagation (RMSprop) technique, and solves the over-fitting problem by using a packet loss technique. The result proves that the prediction performance of the model is superior to that of a support vector machine model, a particle filter model and a simple RNN model. Yu et al propose a new network model averaging method to handle network model uncertainty in RUL prediction of non-lifecycle labeled datasets, verifying the effectiveness of the proposed network model averaging method on Bi-directional LSTM (Bi-L-STM) networks proposed by Huang et al that simultaneously learn forward and backward time related information. Ren et al combines a self-encoder with a Deep Neural Network (DNN) for RUL prediction for lithium ion batteries. Shen et al utilize Deep Convolutional Neural Networks (DCNN) to estimate the capacity of the lithium ion battery on-line, which has higher accuracy and robustness in capacity estimation.

Hybrid-based approaches primarily combine two or more model-based or data-driven approaches to improve prediction performance.

In the last few years, the discovery of the relationship between monitored system data and the corresponding RUL using data-driven methods has received increasing attention. A number of researchers have proposed neural network-based methods to improve RUL prediction accuracy, especially deep learning network structures with great potential in PHM and RUL estimation. Although the structure of the deep network is good, the hidden layers are many, a large number of connection parameters exist between layers, and the structure is very complex, so that a large amount of time is consumed in the network training process, and a large amount of memory is occupied. In addition, designing the network and adjusting the parameters is also an extremely large challenge. It is becoming desirable for researchers to find a way to accurately predict RUL while providing cost savings.

The ELM is a special single hidden layer feedforward neural network, which only has one hidden layer, has simple network structure, does not need back propagation technology to update parameter values, has high learning speed and strong generalization capability, can be used for classification, prediction and the like, and is widely applied to the fields of computer vision, bioinformatics, environmental science and the like. Ginger and the like extract equal pressure drop discharge time as an indirect life characteristic parameter of the lithium battery, and an ELM is used for constructing a relation model of the parameter and the battery capacity, so that the RUL of the lithium battery is indirectly predicted. Wu et al introduced a firefly (GSO) algorithm to optimize the input weight and hidden layer bias value randomly generated by ELM, and adopted the isobaric pressure drop discharge time sequence to perform indirect RUL prediction, thus reducing the prediction error of the algorithm. Razavi-Far et al use one-step and multi-step advance prediction from an extreme learning machine as a prediction module, and estimate RUL using constant-current experimental capacity data using a variety of prediction strategies. Ma et al introduced the idea of generalized learning (BL), developed generalized learning-extreme learning machine (BL-ELM), and this method constantly expands the input layer, increases the nodes of input layer, has greatly improved the ability of the effective characteristic information in the network capture data.

The ELM algorithm first gives a learning sample, the number of hidden nodes and the type of activation function, and then randomly generates weights between the input layer and the hidden layer and biases for neurons of the hidden layer. In general, the weights are evenly distributed between [ -1,1], and the offsets are evenly distributed between [0,1 ]. After all initial parameter values are set, a hidden layer output matrix can be obtained through calculation, and the optimal output weight is calculated by minimizing the square error of a training sample. So far, the learning algorithm of the ELM is completed, and the optimal parameters and the minimum training error can be obtained. In order to carry out subsequent result prediction, the input matrix of the test set is input into the first-layer full-link network of the ELM, and the obtained result is multiplied by the output weight beta to obtain a final prediction result. The structure is shown in fig. 1.

It can be easily found by analyzing and understanding the studies of the former scholars that: when the traditional ELM is used for predicting the RUL of the lithium battery, a plurality of commonly used activation functions such as an S-type function, a Gaussian function, a polynomial function, a hard limit function, a sine function and the like cannot be well predicted, when the relu function which is very popular at present is introduced as the activation function, the RUL can be predicted at some time, uncertain factors exist still, the two problems of result divergence and low precision can occur, and the robustness and the generalization capability of the algorithm are greatly weakened. The prior ELM improvement mainly starts from the aspects of continuously increasing or reducing the number of nodes of a hidden layer according to a certain rule, optimizing a parameter algorithm and introducing a kernel function, an input layer and the hidden layer of the ELM are all fully connected all the time, and an ELM structure with local connection does not get wide attention. At the same time, it is not difficult to find that these improvements increase the amount of calculation to some extent, and there are still many parameters to be initialized, for example, if the selection is not proper, the predicted result will be affected to some extent. In addition, the single-layer convolution square pool architecture with random weights proposed by Andrew m.saxe et al and the Local Receptive field (ELM-LRF) of ELM proposed by Huang et al both demonstrate that constructing a specific network structure can greatly improve system performance, exhibit better effects than some deep Learning algorithms, and greatly improve training speed. Lin et al propose to replace the full link layer with the global average pooling method, i.e., perform global average pooling on each feature map obtained from the last multi-layer sensor convolutional layer, so that each feature map can obtain an output, thereby saving parameters, reducing the complexity of the network, and avoiding the phenomenon of over-fitting.

Disclosure of Invention

The invention aims to provide a lithium ion battery life prediction method based on an improved ELM (element-free mass spectrometer), and two classical methods, namely a convolutional neural network and a global average pooling method, are introduced to improve the ELM in two different ways. First, in order to solve the problem that the locally connected ELM structure is not paid extensive attention, the first improved algorithm is to change the fully-connected relation between the input layer and the hidden layer into convolution and pooling. Considering that in the traditional ELM algorithm and the improved implementation thereof, the nodes of the input layer are all in a complete connection relationship with the nodes of the hidden layer and the local correlation can not be learned, and the CNN is used as a special deep feedforward neural network, the convolution kernel parameters of the CNN are shared, and the layers are in a sparse local connection form, so that the parameter redundancy caused by complete connection is greatly reduced, therefore, the complete connection relationship between the input layer and the hidden layer of the ELM is changed into convolution and pooling operation, and the subsequent mode of acquiring the output weight matrix beta is consistent with the traditional ELM. Secondly, because the values of the convolution kernel and the bias in the first improved algorithm are randomly generated according to a common value-taking method in the traditional ELM, the result of each operation has uncertainty, a good-time and bad-time prediction result is easy to occur, and the algorithm robustness is poor. Based on the above reasons, inspired by the idea of global average pooling, the second improved algorithm of the present invention is to directly change the full connection relationship between the input layer and the hidden layer of the ELM into pooling, and similarly, the output weight matrix β is obtained as in the standard ELM. This method eliminates the uncertainty of the prediction result, since there is no random generation of values and arbitrary selection of activation functions. It should be noted that, since the data of the input layer is only a matrix, i.e. it is equivalent to a characteristic map, the traditional average pooling method is used to reduce the dimension of the matrix instead of directly performing the global average pooling on the matrix. The invention verifies the effectiveness of the improved ELM algorithm by adopting two groups of lithium battery experimental data provided by the advanced life cycle engineering center of Maryland university and the Emms prediction excellence center of the American national aerospace office.

The invention relates to a lithium ion battery service life prediction method based on improved ELM, which adopts the following technical scheme for solving the problems: the first improved algorithm is that firstly the full connection relation between the input layer and the hidden layer of the ELM is changed into convolution operation, i.e. a convolution kernel with a conventional size is introduced, the convolution kernel is regarded as the deformation of the original hidden layer weight and is convolved with the data of the input layer, and then the extracted characteristic matrix passes through an averaging poolTransforming to obtain a hidden layer output matrix H, and transforming Moore-Penrose generalized inverse H of H ⁺ And multiplying the output weight matrix beta by the output matrix of the training set to complete the learning process of the algorithm, and finally carrying in the data of the test set to predict the result. The second improved algorithm is to change the full connection relation between the input layer and the hidden layer of the ELM into pooling, that is, the data of the input layer is directly subjected to pooling operation to obtain the hidden layer output matrix H, and similarly, moore-Penrose generalized inverse H of H ⁺ And multiplying the output weight matrix beta by the output matrix of the training set to complete the learning process of the algorithm, and finally carrying in the data of the test set to predict the result.

The invention relates to a lithium ion battery service life prediction method based on improved ELM, which specifically comprises the following steps:

step101, preprocessing data;

step102. Initialize parameters;

step103, performing convolution operation on the input matrix of the training set to obtain a feature matrix tempH of the training set;

step104, calculating a nonlinear characteristic output matrix tH and a hidden layer output matrix H of the training set through the characteristic matrix tempH extracted in Step 103;

step105, calculating Moore-Penrose generalized inverse H of the hidden layer output matrix H of the training set obtained by Step104 ⁺ ；

Step106, calculating an output weight matrix beta; the output weight matrix beta passes through a training set output matrix T _ train and a generalized inverse H ⁺ The multiplication results in that: beta = H ⁺ T_train。

Step107. Calculating the characteristic matrix of the test set by using the test set input matrix P _ test obtained in Step101

Step108. Feature matrix extracted by Step107

Computing a non-linear feature output matrix for a test set

And hidden layer output matrix

Step109. Hidden layer output matrix of test set obtained by Step108

Multiplying the output weight matrix beta obtained by Step106 to obtain a final prediction result, namely:

the Step101 data preprocessing specifically comprises the following steps: starting from a cycle period k =1, taking data of every 10 continuous cycle periods as a group of input, taking data of the next cycle period as output, respectively constructing two groups of lithium battery data in such a way, and selecting a prediction starting point T after construction is completed to obtain P _ train; t _ train; p _ test; t _ test, wherein P _ train is a training set input matrix, and T _ train is a training set output matrix; p _ test is a test set input matrix; t _ test is the test set output matrix.

The initialized parameters of Step102 comprise: initializing IW; b, where IW is the convolution kernel and B is the offset.

Wherein Step103 is specifically as follows: and taking the input matrix of the training set preprocessed by Step101 as data of an input layer, and performing convolution operation on the input matrix and a convolution kernel to obtain a feature matrix tempH of the training set, wherein the tempH = P _ train _ IW + BiasMatrix, and the BiasMatrix is a copy expansion matrix of the bias B.

Wherein Step104 is specifically as follows: the nonlinear characteristic output matrix tH is obtained by inputting the characteristic matrix tempH of the training set into an activation function for nonlinear characteristic mapping, namely: tH = g (tempH), where g (x) is the activation function. And obtaining a hidden layer output matrix H of the training set by average pooling of tH.

Wherein Step107 is specifically as follows: by the same convolution as in Step103Convolution operation is carried out on the input matrix of the test set through kernel and bias to obtain the feature matrix of the test set

The specific steps are the same as Step103.

Wherein, step108 obtains the nonlinear characteristic output matrix of the test set

And hidden layer output matrix

The concrete steps are the same as Step104.

Further, the present invention provides a lithium ion battery life prediction method based on an improved ELM, which specifically includes the following steps:

step201, preprocessing data;

step202, calculating a hidden layer output matrix H of the training set;

step203. Calculating Moore-Penrose generalized inverse H of the hidden layer output matrix H of the training set obtained by Step202 ⁺ ；

Step204, calculating an output weight matrix beta; outputting matrix T _ train and generalized inverse H through training set ⁺ The multiplication results in that: beta = H ⁺ T_train；

Step205, calculating a hidden layer output matrix of the test set through the test set input matrix P _ test obtained at Step201

Step206. Obtain the final prediction result Y, and obtain the hidden layer output matrix of the test set through Step205

Multiplying the output weight matrix beta obtained at Step204 to obtain a final prediction result, namely:

the Step201 data preprocessing specifically comprises the following steps: starting from a cycle period k =1, taking data of every 10 continuous cycle periods as a group of input, taking data of the next cycle period as output, respectively constructing two groups of lithium battery data in such a way, and selecting a prediction starting point T after construction is completed to obtain P _ train; t _ train; p _ test; t _ test, wherein P _ train is a training set input matrix, and T _ train is a training set output matrix; p _ test is a test set input matrix; t _ test is the test set output matrix.

Wherein Step202 is as follows: and taking the training set input matrix P _ train preprocessed by Step201 as data of an input layer, and directly carrying out average pooling on the hidden layer output matrix H of the training set by the training set input matrix P _ train to obtain the hidden layer output matrix H.

Wherein, the hidden layer output matrix of the test set obtained by Step205

The test set input matrix P _ test is pooled in the same manner as in Step202, and the specific steps are the same as in Step202.

The lithium ion battery life prediction method based on the improved ELM has the advantages and effects that: the first improved algorithm introduces the idea of local perception visual field, changes the full connection relation between the input layer and the hidden layer, reduces parameter redundancy and greatly reduces the calculated amount, so that the algorithm prediction result is more accurate than that of the traditional ELM. The second improved algorithm is carried out on the basis of the first improved algorithm, the operation of convolution is omitted, the input layer data is directly subjected to pooling, and due to the fact that parameter initialization and selection of the type of an activation function are omitted, the prediction result of the algorithm is more stable, and the robustness is stronger.

Drawings

FIG. 1 is a flow chart of a conventional ELM algorithm

FIG. 2a is a flow chart of the improved algorithm one (embodiment 1) of the present invention

FIG. 2b is a flow chart of the improved algorithm two (embodiment 2) of the present invention

FIG. 3a is a graph showing the capacity variation of the data A3, A5, A8 and A12 of the lithium ion batteries of 4 groups of the advanced Lifecycle engineering center of Maryland university

FIG. 3B is a graph showing the capacity variation of the data B5, B6, B7 and B18 of 4 groups of lithium ion batteries of Emms prediction of the national aerospace agency

FIG. 4a is a graph comparing the true value of A12 battery data with the results of three algorithms

FIG. 4b is a graph showing the absolute error of the true A12 cell data and the absolute error of the three algorithms

FIG. 5a is a graph comparing the real value of the B5 battery data with the results of the three algorithms

FIG. 5B is a diagram showing the absolute error between the true value of the B5 battery data and the three algorithms

FIG. 6a is a graph showing the RMSE for A12 battery data

FIG. 6b shows a MAPE plot of A12 battery data

FIG. 7a is a graph of RMSE for B5 battery data

FIG. 7B shows a MAPE plot of B5 battery data

FIG. 8a shows the RMSE values of the algorithms under different lithium battery data of the CALCE group

FIG. 8b shows the RMSE values of the algorithms under different lithium battery data of NASA group

FIG. 8c shows MAPE values of algorithms of different lithium battery data of CALCE group

FIG. 8d shows MAPE values of algorithms under different lithium battery data of NASA group

FIG. 8e shows the RUL values of algorithms of CALCE set under different lithium battery data

FIG. 8f shows the RUL values of the algorithms under different lithium battery data of NASA group

Detailed Description

The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.

Example 1

The invention relates to a lithium ion battery service life prediction method based on improved ELM, which comprises the following specific processes:

as shown in fig. 2a, a first improved algorithm:

step one, data preprocessing. The method specifically comprises the following steps: starting from a cycle period k =1, taking data of every 10 continuous cycle periods as a group of input, taking data of the next cycle period as output, respectively constructing two groups of lithium battery data in such a way, and selecting a prediction starting point T after construction is completed to obtain P _ train; t _ train; p _ test; t _ test, wherein P _ train is a training set input matrix, and T _ train is a training set output matrix; p _ test is a test set input matrix; t _ test is the test set output matrix.

Step two, initializing parameters, including: initializing IW; b, where IW is the convolution kernel and B is the offset.

And step three, performing convolution operation on the input matrix of the training set to obtain a feature matrix tempH of the training set. The method specifically comprises the following steps: and (2) taking the input matrix of the training set preprocessed in the first step as data of an input layer, and performing convolution operation on the input matrix and a convolution kernel to obtain a feature matrix tempH of the training set, wherein the tempH = P _ train _ IW + BiasMatrix, and the BiasMatrix is a copy expansion matrix of the bias B.

Step four, calculating a nonlinear characteristic output matrix and a hidden layer output matrix tH of the training set through the characteristic matrix tempH extracted in the step three; H. the method specifically comprises the following steps: the nonlinear feature output matrix tH is obtained by inputting the feature matrix tempH of the training set into an activation function for nonlinear feature mapping, namely: tH = g (tempH), where g (x) is the activation function. And obtaining a hidden layer output matrix H of the training set through average pooling by tH.

Step five, calculating Moore-Penrose generalized inverse H of the training set hidden layer output matrix H obtained in step four ⁺ ；

And step six, calculating an output weight matrix beta. Using training set output matrix T _ train and generalized inverse H ⁺ The multiplication results in an output weight matrix β, i.e.: beta = H ⁺ T_train。

Step seven, using the test set input matrix P _ test obtained in the step one to calculate the characteristic matrix of the test set

The method specifically comprises the following steps: using the same convolution kernel in step threeAnd performing convolution operation on the input matrix of the test set by the bias to obtain the feature matrix of the test set

Namely:

wherein the values of IW and B are the same as those of step three.

Step eight, extracting the feature matrix through the step seven

Calculating the nonlinear characteristic output matrix and the hidden layer output matrix of the test set

The method specifically comprises the following steps: the same as the fourth step, the feature matrix tempH of the training set is input into the activation function for nonlinear feature mapping to obtain a nonlinear feature output matrix

Namely:

wherein g (x) is an activation function. Hidden layer output matrix of training set

By

Obtained by average pooling.

And step nine, obtaining a final prediction result Y. The method specifically comprises the following steps: hidden layer output matrix of test set obtained through step eight

And multiplying the output weight matrix beta obtained in the step six to obtain a final prediction result, namely:

step ten, evaluating the algorithm.

Example 2

As shown in fig. 2b, a second improved algorithm:

step one, data preprocessing. The method comprises the following specific steps: starting from a cycle period k =1, taking data of every 10 continuous cycle periods as a group of input, taking data of the next cycle period as output, respectively constructing two groups of lithium battery data in such a way, and selecting a prediction starting point T after construction is completed to obtain P _ train; t _ train; p _ test; t _ test, wherein P _ train is a training set input matrix, and T _ train is a training set output matrix; p _ test is a test set input matrix; t _ test is the test set output matrix.

And step two, calculating a hidden layer output matrix H of the training set. The method specifically comprises the following steps: and (3) using the training set input matrix P _ train preprocessed in the step one as data of an input layer, and directly carrying out average pooling on a hidden layer output matrix H of the training set by using the training set input matrix P _ train to obtain the hidden layer output matrix H.

Step three, calculating Moore-Penrose generalized inverse H of the training set hidden layer output matrix H obtained in the step two ⁺ 。

And step four, calculating an output weight matrix beta. Using training set output matrix T _ train and generalized inverse H ⁺ The multiplication results in an output weight matrix β, i.e.: beta = H ⁺ T_train。

Step five, calculating a hidden layer output matrix of the test set through the test set input matrix P _ test obtained in the step one

The method specifically comprises the following steps: hidden layer output matrix of test set

The test set input matrix P _ test is averaged and pooled in the same manner as in step two.

And step six, obtaining a final prediction result Y. The method specifically comprises the following steps: hidden layer output matrix of test set obtained through step five

And multiplying the output weight matrix beta obtained in the step four to obtain a final prediction result, namely:

and seventhly, evaluating an algorithm.

The experiment is simulated by using MATLAB R2020B software, based on lithium battery experimental data of advanced life cycle engineering center of Maryland university and Ames prediction and excellence center of national aerospace Bureau, A3, A5, A8, A12 and B5, B6, B7 and B18 are selected as experimental data in the experiment, and the two groups of lithium ion battery experimental data are respectively shown in FIG. 3a and FIG. 3B.

The lithium battery supplied by the CALCE center is 4 lithium batteries of the same type from the same manufacturer, and both have a graphitic carbon anode and a lithium cobalt oxide cathode. Under the condition of room temperature, an Arbin BT2000 battery test system is adopted to carry out a plurality of charge-discharge tests, and when the charge or discharge voltage reaches a specified cut-off voltage, a charge or discharge process is considered to be completed. The rated capacity of the battery was 0.9Ah, and the constant discharge current was 0.45A.

The lithium battery data provided by NASA Ames PCoE was obtained from a second generation 18650 size lithium ion battery produced by the national laboratory of idaho, using 3 different (charging, discharging and impedance) operating operations at room temperature (24 ℃). Charging was performed in Constant Current (CC) mode of 1.5A until the battery voltage reached 4.2V, and then charging was continued in Constant Voltage (CV) mode until the charging current dropped to 20mA. The discharge was run in constant current mode of 2A until the voltage of batteries B5, B6, B7, B18 dropped to 2.7V, 2.5V, 2.2V and 2.5V, respectively. Impedance measurements were made by Electrochemical Impedance Spectroscopy (EIS) frequency scanning from 0.1Hz to 5 kHz. Repeated charge and discharge cycles result in accelerated battery degradation, while impedance measurements provide insight into how internal battery parameters vary. Although the cycle discharge conditions of these four batteries were slightly different, they were considered to have a rated capacity of 2Ah.

From both fig. 3a and fig. 3b, it can be seen that the storage capacity of the lithium battery gradually decreases with the increase of the charging and discharging times, but when the lithium battery is in a static state during the charging and discharging processes, the reaction products are dissipated, so that the electrochemical performance of the lithium battery is relatively recovered relative to the previous cycle period during the degradation process. It is apparent that the available capacity of the next cycle increases, and thus is referred to as a capacity regeneration phenomenon of the lithium battery. Their capacities have different degradation rates for the same type of battery. In the figure, the horizontal axis represents the number of charge and discharge cycles, and the vertical axis represents the actual capacity value of the lithium battery after each charge and discharge. In addition, it is generally considered that the end of life (EOL) is indicated when the discharge capacity of the lithium battery decreases to 60% to 80% of its rated capacity. Therefore, since 80% of the rated capacity of the first group of cells is set as the end point of life, the failure threshold value shown by the purple dotted line in fig. 3a is 0.72Ah, and similarly, since 71% of the rated capacity of the second group of cells is set as the end point of life, the failure threshold value shown by the dotted line in fig. 3b is 1.42Ah. This gives: the end-of-life (i.e., the real value of RUL) of the first group of lithium batteries A3, A5, A8, a12 is: 47. 188, 131, 208; the end-of-life (i.e., the real value of RUL) for the second group of lithium batteries B5, B6, B7, B18 is: 115. 105, 159, 89.

In the experiment, the traditional ELM algorithm and two improved ELM algorithms are respectively used for training and predicting sample data. The number of hidden layer nodes of the traditional ELM is equal to the number of samples of a training set, the weight w between an input layer and a hidden layer and the bias b of the neurons of the hidden layer are randomly and uniformly generated, and a sine function is adopted as an activation function. In a first algorithm of the improvement, the convolution part is set: the size of the convolution kernel is 3 multiplied by 3, the values of the convolution kernel and the offset are initialized randomly and uniformly, the step length is 1, the mode is all-zero filling, and the activation function also adopts a sine function; the pooling part is set in such a way that the size of the pooling core is 2 x 1, the step size is 1, and the pooling is averaged. At the same time, the arrangement of the pooling part in the modified second algorithm is the same as that of the first method. Unlike the pooling kernels having the same length and width that are usually used, the pooling kernel having the size of 2 × 1 is used, and this is taken to be able to calculate the output weight matrix β later. After the training set and the test set are divided and the prediction starting point is selected, the size of the output matrix Y of the training data is fixed, and in order to obtain the value of beta, the size of the hidden layer output matrix H obtained after pooling operation is matched with the size of Y, so that the calculation can be continued.

Example 1, a specific process of a first improved algorithm is as follows:

1. normalizing the divided training set and the test set to obtain:

Tn_train＝[1.00000.95180.78770.7840…-0.7051-1.00000.7966-0.8837]

Tn_test＝[-1.0227-0.9287-1.0667-1.0456…-34.3215-34.5820-34.6421-34.6118]

T_test＝[0.86340.86560.86240.8629…0.07920.07310.07170.0724]

2.

3.

4.

5.

6.

7.

8.

9. the inverse normalization is carried out to obtain a predicted value,

example 2, the specific process of the second modified algorithm is as follows:

1. normalizing the divided training set and the test set to obtain:

Tn_train＝[1.00000.95180.78770.7840…-0.7051-1.00000.7966-0.8837]

Tn_test＝[-1.0227-0.9287-1.0667-1.0456…-34.3215-34.5820-34.6421-34.6118]

T_test＝[0.86340.86560.86240.8629…0.07920.07310.07170.0724]

2.

3.

4.

5.

6. and (4) inverse normalization is carried out to obtain a predicted value:

the effectiveness of the two improved ELM algorithms of the present invention is verified by using the lithium battery experimental data of the advanced Life cycle engineering center of Maryland university and the national aviation and space agency Emms prediction excellence center, and compared with the conventional ELM algorithm.

The experiment is compared with the traditional ELM algorithm in order to show the accuracy of the prediction effect of the improved ELM algorithm.

Fig. 4 and 5 correspond to battery data a12 and B5, fig. 4a and 4B show the prediction result and absolute error of a12 when the prediction start point is T =80, and fig. 5a and 5b show the prediction result and absolute error of B5 when the prediction start point is T = 70.

In the figure, the curved solid line represents the true value; the curved dashed line represents the predicted result of ELM; the dotted line represents the predicted result of the first modified algorithm; the point-symbol curve represents the predicted result of the second improved algorithm; the horizontal dashed line is the battery capacity failure threshold; the vertical solid line is the prediction starting point. The prediction effect is significantly improved with the improvement of the algorithm, and as can be seen from fig. 4a and 5a, the line represented by the improved two algorithms is closer to the line represented by the true value than the line represented by the conventional ELM algorithm. Absolute error it can be seen in fig. 4b and 5b that the absolute error of both improved algorithms is small, and the second improved algorithm is the most stable and robust one.

The experiment was completed for 101 cycles, and the resulting RMSE and MAPE after each cycle were recorded and plotted as line graphs as shown in fig. 6 a-7 b. The closer the values of RMSE and MAPE are to 0, the more accurate the prediction method is. It can be seen from the figure that, for the different data sets a12 and B5 and the different initial parameter values, the predicted values of the two improved algorithms have small RMSE and MAPE compared with the conventional ELM algorithm, and particularly, the predicted result of the second improved algorithm is the most stable, and the result obtained by each loop is not changed and does not fluctuate. Therefore, the improved two ELM algorithms of the invention have better prediction performance compared with the traditional ELM algorithm.

As can be seen from fig. 8a to 8f, the prediction effect of the conventional ELM algorithm on the eight battery data sets is not ideal, and even the failure threshold point is not reached sometimes. The prediction effect of the two improved ELM algorithms under different data sets is better than that of the traditional ELM algorithms, and the two improved algorithms not only have smaller values of RMSE and MAPE but also have closer predicted RUL values to the RUL true values along with the increase of the prediction starting points. The first improved ELM algorithm improves the large-amplitude filtering divergence effect of the traditional ELM algorithm, so that the error is reduced, and the second improved ELM algorithm is continuously corrected on the basis of the first improved algorithm because even if the error of the first improved algorithm greatly improves the accuracy of the prediction result, the first improved algorithm still has uncertainty, and the uncertainty causes the prediction result to be possibly accurate in prediction and not ideal in prediction. Therefore, based on this point, the second improved ELM algorithm is generated, and the prediction result is the most stable and more robust.

Claims

1. A lithium ion battery life prediction method based on improved ELM is characterized in that: the method comprises changing the full connection relation between the input layer and the hidden layer of ELM into pooling, i.e. directly pooling the data of the input layer to obtain hidden dataA reservoir output matrix H; and then the Moore-Penrose generalized inverse H of the matrix H ⁺ Multiplying the output matrix of the training set by the output matrix of the training set to obtain an output weight matrix beta, and finishing the learning process of the algorithm; and finally, carrying out result prediction by taking in test set data.

2. The improved ELM-based lithium ion battery life prediction method of claim 1, wherein: the method comprises the following specific steps:

step201, preprocessing data;

step202, calculating a hidden layer output matrix H of the training set;

Step204. Calculate the output weight matrix β: outputting matrix T _ train and generalized inverse H through training set ⁺ The multiplication results in that: beta = H ⁺ T_train；

Step206. Get final prediction result Y: hidden layer output matrix of test set obtained through Step205

3. the improved ELM-based lithium ion battery life prediction method of claim 1, wherein: the Step201 data preprocessing specifically comprises the following steps: starting from a cycle period k =1, taking data of every 10 continuous cycle periods as a group of input, taking data of the next cycle period as output, respectively constructing two groups of lithium battery data in such a way, and selecting a prediction starting point T after the construction is finished to obtain P _ train; t _ train; p _ test; t _ test, wherein P _ train is a training set input matrix, and T _ train is a training set output matrix; p _ test is a test set input matrix; t _ test is the test set output matrix.

4. The improved ELM-based lithium ion battery life prediction method of claim 1, wherein: the specific process of Step202 is as follows: and taking the training set input matrix P _ train preprocessed by Step201 as data of an input layer, and directly carrying out average pooling on the hidden layer output matrix H of the training set by the training set input matrix P _ train to obtain the hidden layer output matrix H.